Assessing the likelihood of drought impact occurrence with extreme gradient boosting: a case study on the public water supply in South Korea

Abstract

Drought is quantified with one or a set of drought indices for monitoring and risk management. These indices have a limited ability to capture drought impacts. Drought impact prediction models have been developed to explore the interactions between the drought impact data and the physical drought indices. This study demonstrates the use of extreme gradient boosting (XGB), a well-known machine learning technique, to predict the likelihood of impact occurrence (LIO) of drought on public water supply as a function of drought indices, with high accuracy and low uncertainty. Using text-based drought impact data from multiple sources, the prediction accuracy of drought LIO on the public water supply of South Korea was evaluated using XGB and reference models (log-logistic, support vector machine, and random forest). We also analyzed receiver operating characteristics and quantified the uncertainty of each model with bootstrapping. This study shows that XGB and random forest have a high level of suitability. However, random forest presents a higher level of uncertainty than XGB for predicting drought LIO on the public water supply in South Korea. Although some limitations exist, the results suggest that text-based drought impact data collected from multiple sources can provide insightful information for drought risk management.